Hello, i’m developing a face recognition model using DS6.2. I was trying to compare the performance for single and multi input and what i have noticed that there is not concolusion about the relationship between FPS and increasing input sources.
1-Input source achieved 116 FPS
2-Input sources achieved 79 FPS
4-Input sources achieved 95 FPS
how can increasing number of sources to 4 gave better results than 2 inout sources?
Hi,
How are you configuring your batch-size in nvstreammux and in the model config? The only thing that I can think of is that the batch size is 4, and the case with 2 sources is introducing stalling time due to synchronization issues.
Thank you @miguel.taylor . Batch size is set to 2 in case of 2 streams and 4 incase of 4.
That’s something i noticed in the gpu load in some seconds of running the pipeline the GPU load in powerGUI shows zero which i couldn’t find justification for.
one more thing is that with single input running the Facedetection on GPU gave better FPS than running it on DLA. However, unexpectedly this is not the same trend for 2 -input sources. with 2-input sources running Face detection model on DLA gave better FPS than running it on GPU. When i have checked the gpu load i found in case of running on GPU 5 seconds out of 11 the gpu has a load of zero while in other 6 seconds the gpu usage is too low. while this is not the case when running on DLA.
Can you please help me to understand the reason
All clues point to a synchronization issue.
How are you running the pipeline (gst-launch-1.0, gstreamer python, deepstream sample…)?
Can you try adding sync=false to your sink element?
I’m simply running the python3 file.py input video.mp4
i dont have the sync element in my config file, can i share the code through msgs so you please advise how can i do so
i dont have the sync element in my config file, can i share the code through msgs so you please advise how can i do so
Sure, that would be useful. I think strategically adding queues and discarding gstreamer synchronization would probably solve your issue.
Thank you
i have used the queues and discard the async and sync in the code but still same issue
There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.
Thanks
Hi,
Please try to replicate the issue by running deepstream-app, and share us the steps. We will set up AGX Orin developer kit and check.